← Back to Sessions

Evaluation of Deep Learning Phase Picking Models

Description: 

Deep learning phase pickers have been shown to outperform traditional models. Deep learning models are trained using large, labeled datasets and then applied to incoming data or new datasets for automated creation of seismic catalogs. However, they suffer from the generalizability problem, i.e., their performance tends to drop when applied to new regions due to differences in regional geological structures, sources, measurement instruments, and environmental effects. In addition, a large, labeled dataset, required to train deep learning models from scratch, may not be available for a region of interest. Previous studies have demonstrated improved performance on a new region when phase picking models are fine-tuned using a labeled dataset corresponding to this new region. In this work, we evaluate and compare three models: a model trained using STEAD data (Model 0), a model trained using INSTANCE data (Model 1), and a model with weights initialized from Model 0 and fine-tuned using INSTANCE data (Model 2). We present our approach for model performance evaluation and metrics used. Specifically, we evaluate model transferability through transfer learning by evaluating how the amount of data used for fine-tuning impacts performance of Model 2. Through an ensemble-based approach, we also perform uncertainty quantification for model predictions.

Session: Network Seismology: Recent Developments, Challenges and Lessons Learned - III

Type: Oral

Date: 5/2/2024

Presentation Time: 08:15 AM (local time)

Presenting Author: Nidhi

Student Presenter: No

Invited Presentation: 

Authors